Block 3: AI as a Source + Documentation Protocol
Context: The Cauvery river rises in Karnataka and flows into Tamil Nadu before reaching the sea. Karnataka is upstream, Tamil Nadu is downstream, and Kerala and Puducherry hold smaller shares. Who gets how much water has been argued over for more than a hundred years, through colonial agreements, a national tribunal, the Supreme Court, and a standing water authority. The argument returns in every dry monsoon year. You do not need to know that history to answer today's question.
Image: two AI responses to the same prompt 'When was the Cauvery dispute resolved?' One says '2018 by Supreme Court verdict.' One says '2007 by the Cauvery Water Disputes Tribunal.' Both are confidently stated.
Task: Which is accurate? How could you find out without trusting either response? Which entry in the Resources Library would you check first? What does this tell you about AI as a source?
Answer in your Field Notebook, dated.
I can compare three AI responses to the same paper, check a specific claim against the paper itself, and produce a complete AI Documentation Template naming at least one specific thing an AI got wrong.
I can find sources, judge them with OPVL, and communicate what I learned honestly.
"Obtaining, Evaluating, and Communicating Information"
Where this runs:
- U0 B1: Welcome + How We Work
- U0 B2: OPVL Across Source Types
- U0 B3: AI as a Source + Documentation Protocol ← you are here
- U0 B4: Source Dossier Launch + Choosing Your Topic
- U0 B5: Source Dossier Work Day 2: AI Source + Documentation
- U2 B9: Landscape Readings Submitted + Class Share-out
Assessed:
Source Dossier (all strands) + every OPVL component after it
On track looks like:
Your Limitations are specific to the source in front of you, never 'it might be biased.'
I can check an AI's answer against real sources and decide: use it, fix it, or bin it.
In brief: "Evaluate whether an AI output should be accepted, revised, or rejected (paraphrased; full framework at ailiteracyframework.org)"
Where this runs:
- U0 B1: Welcome + How We Work
- U0 B2: OPVL Across Source Types
- U0 B3: AI as a Source + Documentation Protocol ← you are here
- U0 B4: Source Dossier Launch + Choosing Your Topic
- U0 B5: Source Dossier Work Day 2: AI Source + Documentation
- U1 B1: Phenomenon Launch: What Makes a Planet Habitable?
- U1 B3: Mars as Calibration: What Does Earth Have That Mars Doesn't?
- U1 B5: Exoplanet Detection: Transits and What They Tell Us
- U1 B6: Reading Exoplanet Data Cards: Choosing Your Planet
- U1 B7: Habitability Framework: Building the Comparison
- U2 B3: Rivers as Landscape Sculptors
- U2 B6: Field Photo / Map Analysis Work Day
- U2 B7: Reading the Landscape: Worked Example + OPVL
- U4 B5: IPCC AR6 Deep Dive + Chennai Climate Brief Assigned
- U4 B6: Climate Models: How They Work, How They're Wrong, How to Read Them
Assessed:
Every AI Documentation Template: the 'one thing the AI got wrong' field
On track looks like:
You catch specific errors ('it said 2024; the report says 2023'), and you've stopped writing 'AI can hallucinate.'
I choose when to use AI and when to think it through myself, and I can defend either choice.
In brief: "Decide whether AI is the right tool for a task before using it (paraphrased)"
Where this runs:
- U0 B3: AI as a Source + Documentation Protocol ← you are here
- U0 B5: Source Dossier Work Day 2: AI Source + Documentation
- U0 B7: Defenses + Unit Close
- U1 B8: Goldilocks Report Drafting: Substantive Writing Block
- U1 B10: NASA Committee Q&A + Unit Reflection
- U2 B8: Landscape Reading Drafting + Peer Feedback
- U4 B10: Climate Brief Work Day + Worked-Example Brief + Peer Feedback
- U4 B11: Climate Brief Final Drafting + Panel Rehearsal
Assessed:
AI-use log across the year; the defended not-using-AI reflection
On track looks like:
You can name a step where you chose NOT to use AI, and your reason survives a 'why?'
What today is about
The move that makes this class different from most.
We open with what happened on your screens. Last night you logged into the class AI platform for the first time and ran the OPVL tool on Lovelock (1965). Bring out your output, what was helpful, and where it was difficult. If you could not get in at all, that is the finding we want first, not something to be embarrassed about.
Then we go further. AI outputs are sources, and they get the same OPVL framework and the same critical eye. By the end of today you will be holding four responses to one paper you have actually read: yours from last night, two that Ms. Jayanthi/Mr. Ignash ran on different models, and one more you run yourself. Same content, different source types, and they will not say the same things. The question you leave with is not which one is nicest to read. It is which one said anything specific enough that you could catch it being wrong. Then we introduce the habit you'll carry for the rest of the year: every time you use AI in your work, you fill out the AI Documentation Template.
The documentation isn't there to police you. It's there to make you notice what the AI did versus what you did. Students who can articulate that distinction learn more from AI than students who can't. Honest disclosure = no penalty. Hidden AI use = academic integrity violation. That's the whole protocol. Want to see the protocol working on a real document? The exemplar dossier's own AI Documentation page is a live example: the exemplar was built with AI and documents itself the same way you will.
One more thing, said plainly: on our class AI platform, Ms. Jayanthi/Mr. Ignash can see student conversations. That's not surveillance of your thinking; it's the same principle as the Documentation Template. In this class, AI use is visible by design. Work as if your process will be read, because it will be.
Before you start the work
Use all of these. The comparison page carries the responses today's work is built on, and the protocol is your reference for the template you submit at the end. Notes go in your Class Notebook as a Power outline.
Taking notes? Power 1 seeds for today
Everyone does this
Today's work
One task today. Everyone runs the same comparison, because the whole point is that we can put our answers side by side.
Ms. Jayanthi/Mr. Ignash has already uploaded Lovelock (1965), the paper you OPVL'd last class, and run the same prompt through two models: ChatGPT 4.1 and Claude 4.8. You will add a third, and you already have a fourth from last night. Then they get checked against the paper. This is this morning's bell ringer again, except this time you are holding the primary source.
- Get set up properly. See the box below for your section's platform. If you got in last night, you are done with this step in thirty seconds. If you did not, this is the fifteen minutes to fix it, and Ms. Jayanthi/Mr. Ignash will come to you.
- Read the two you were given. The ChatGPT 4.1 and Claude 4.8 responses are on the Two models, one paper page, in full. Do not skim them. You are looking for where they differ, and the page is searchable for a reason.
- Add a third. Download the Lovelock PDF from Google Classroom, upload it to a new conversation, and pick a model nobody demonstrated. Run this prompt, word for word, so your output is comparable to the other two:
"Summarise this paper. Then write an OPVL analysis of it: Origin, Purpose, Value, Limitation." - Compare at the claim level. Pull three specific claims from each response, including the one you ran at home last night. The same claim in different words is not a disagreement. Mark only where they genuinely differ, and mark what one says that the others leave out entirely.
- Find something checkable and check it. Pick a claim specific enough to be verified: a date, a journal, a name, a chemical, a quoted phrase. Check it in the PDF, not the interactive reader. The reader is shortened for reading support, so something missing from it may still be in the paper. Does the paper actually say that?
- Document it. Full OPVL on your third response, the Limitation line on the others, and one AI Documentation Template covering all of it.
Check Google Classroom for which platform your section uses. Sign in with your school account. This is the tool you will use for the rest of the year, so it is worth getting properly set up today rather than half-working.
On BoodleBox: open a new chat and upload the PDF there. For step 3, choose any model other than ChatGPT 4.1 or Claude 4.8 and run the prompt on that same uploaded paper. You are changing the tool and holding everything else constant.
On Flint: you cannot pick the model, so for step 3 run the prompt in a genuinely new conversation and upload the paper again. Do not type "do it again" in the same thread, because it will read its own first answer and agree with itself. You are changing nothing at all and seeing what happens anyway. If your run comes back nearly identical to itself, that is your result, and step 5 still applies.
These responses do not contradict each other the way the two Cauvery answers did. Nothing here says "2007" against "2018". They differ in how specific they are willing to be. One names the journal, the volume, the grant number, and the exact chemicals. One says "the 1960s" and "several experimental approaches."
So the question is not just "which one disagrees." It is: which one said enough to be caught being wrong? A response too vague to be checked is not a safer response. It is an unfalsifiable one, and that belongs in your Limitation line.
That is a real result, not a failed task. Write it down, then do step 5 anyway, because three responses agreeing tells you nothing about whether any of them is right. If they agree and they are all wrong about the same thing, you have found the most important idea in this block: consistency is not accuracy.
Put today's OPVL beside your own OPVL of the paper from Block 2. What specifically is different in the Limitation line? Reading a paper and reading a summary of a paper are different acts, and the Limitation line is where that difference shows up.
Work with a partner and open the sample AI Documentation entry from Google Classroom. It has some fields done well and some done badly. Use it as the model for what "specific" looks like, then come back to your own.
Open your Class Notebook and type today's entry header as Heading 2:
A: Aug 14 | U0 B3 | AI as a Source + Documentation ProtocolB: Aug 17 | U0 B3 | AI as a Source + Documentation Protocol
The Source Evaluator marks where you are: [O 1/4] for Origin through to [L 4/4] for Limitations. It will move you to the next letter after about three exchanges, even when your answer still feels rough. That is deliberate. Last time this ran, every conversation spent the whole period on Origin and never reached Limitations, which is the letter your Dossier is weakest on and the one the rubric looks hardest at. A rough L beats a flawless O.
What you're submitting today
This is the practice run. The Source Dossier (due at the start of Block 7) will require this same template, with the same specificity. Today's the time to get the habit right.
Full template, all seven fields, covering the three responses you worked with on Lovelock (1965): the two you were given and the one you ran yourself. The "one thing the AI got wrong" field must be specific, and today you have the paper itself in front of you, so "I couldn't find anything" is not available. Open the template and its guiding questions if you need the field-by-field prompts. Lovelock returns in Unit 1 Block 1; your OPVL + AI Doc work today is the bridge.
Submit via Google Classroom (link posted in August)One question before you leave
Three to five minutes. Your answer is saved to your reflection journal, where you can read back everything you have written this year.
No grade. Surface the strangest one to the whole class.
Today's reflection is in Google Classroom, under Reflection Journal.
Link posted in AugustSign in with your school account. Your teacher can see what you write here.